Archive
推論 runtime、量化、 serving、KV cache、成本與效能。
推論系統
Based on host memory and PCIe bandwidth, FreeToken dynamically decides whether experts missing from the cache should be transferred to the GPU or computed directly on the CPU. The paper reports 1.3× to 2.1× higher decoding throughput than the strongest baseline across five consumer-grade systems, but the linked code repository is not currently publicly accessible.
推論系統
DriveCache uses vehicle translation and rotation known before generation to schedule diffusion-model feature reuse across different driving scenarios. On Wan2.2 A14B, it delivers 1.276 dB higher PSNR than SeaCache at a nearly identical acceleration level, though the code has not yet been released.
AI 安全
A new study uses an additive log-odds model to measure subtle biases introduced by paraphrases, typos, animal names, and JSON fields, then combines cues that push in the same direction into a prompt. Some optimized prompts transfer across models, though the current experiments focus on artificially designed binary-choice questions.
代理工具
UCL researchers modeled agents, files, messages, and read/write events as a temporal network and analyzed 1,902 controlled programming tasks. The results show that collaboration topology is determined mainly by task structure; merely designating a coordinator in the prompt did not create an actual hub.
AI 安全
A new study finds that several guard models and activation probes show almost no change in judgment accuracy when rules are removed, shuffled, or replaced. In cross-rule tests designed to eliminate textual shortcuts, fast detectors perform near chance; only models using step-by-step reasoning clearly recover the ability to combine rules with scenarios.
推論系統
ParaTempo periodically probes the provisional answer distributions of individual reasoning paths, then independently prunes, retires early, or forks them. On four benchmarks, Qwen3.5-35B-A3B averaged 1.1 percentage points lower accuracy than fixed 16-path self-consistency, while reducing latency by 21.8% and generated tokens by 30.3%.
推論系統
The new release stops overriding GPU memory information with the system’s `MemAvailable` value in Linux HIP builds and instead trusts `hipMemGetInfo`. The fix targets AMD APUs with configurable UMA carveouts and reduces the risk of the model loader overestimating capacity and failing later.
訓練基礎設施
Rollplex executes video encoding and prompt prefill for the reference model and training actor early, during autoregressive decoding, while preserving synchronous on-policy semantics. The system uses CUDA VMM to manage memory across stages and enables training and inference processes with different tensor-parallel configurations to share compatible weights.
CForce 沿用擴散語言模型自己的生成軌跡，把後期較完整的預測蒸餾回早期遮罩狀態。LLaDA2.1-mini 在提高平行解碼量的同時維持準確度，但證據目前集中於同一模型家族與作者控制的評測。
推論系統
The new release configures threads and work-groups according to quantization block size, preventing over- or undersubscription in the SYCL copy kernel. Microbenchmark throughput on Arc 70 rose from 20.21 to 158.19 GB/s, but this does not represent the speedup in token generation for a complete model.
最新模型
Intern-S2-Mobius separates the knowledge storage and attention-based reasoning traditionally tied to each Transformer layer, allowing multiple Reasoners to repeatedly query the same global Memory. The open weights can be loaded with LMDeploy, vLLM, or Transformers, but the nearly 4× acceleration remains a result from the official evaluation under specific conditions.
最新模型
Qwen has released the weights for Qwen3.8-27B, using a 64-layer hybrid Gated DeltaNet/Gated Attention architecture to reduce long-context costs while providing native visual understanding and adjustable reasoning depth. Official agent benchmarks show substantial gains, but some datasets, corrected tasks, and execution frameworks were configured by Qwen, so the results should not be treated as a fair cross-vendor ranking.